Military Fine-tuning Dataset Construction Method for Chinese Large Language Model

Large language models (LLMs) have seen increasing utilization in various general fields. However, their application in the Chinese military domain has not yielded satisfactory results. A significant factor is that due to particularity of the military domain, acquisition of high-quality large-scale Chinese datasets is not only costly, but also extremely difficult during implement. Recently, a turning point has been reached in the construction of fine-tuning datasets for military-domain Chinese LLMs, that is, with the iteration of open-source Chinese LLMs, their reasoning capabilities have notably improved, enhancing the reliability of generated data. Besides, data generation methods based on LLMs are constantly emerging. The approach of leveraging the knowledge base and reasoning abilities of LLMs to generate datasets, subsequently fine-tuning the large models, has been validated multiple times.The motivation of this work is to explore the application of LLMs to generate large-scale fine-tuned datasets for the military domain. Furthermore, enhance the military knowledge base and reasoning capabilities of LLMs. In this work, we proposed three military data generation methods based on LLMs and conducted experiments on two of them. The experimental results demonstrate the efficacy of the approach in generating large-scale military datasets and subsequently, fine-tuning large models. Additionally, we delved into the domain-transfer capabilities of LLMs, providing deeper insights into their potential and limitations.